Using noun class information to model selectional preferences for translating prepositions in SMT

Using Noun Class Information to Model Selectional Preferences for Translating Prepositions in SMT Marion Weller1,2 wellermn@ims.uni-stuttgart.de Sabine Schulte im Walde1 schulte@ims.uni-stuttgart.de Alexander Fraser2 fraser@cis.uni-muenchen.de 1IMS, Universit¨at Stuttgart, Germany 2CIS, Ludwig-Maximilians-Universit¨at M¨unchen, Germany Abstract Translating prepositions is a difficult and under-studied problem in SMT. We present a novel method to improve the translation of prepositions by using noun classes to model their selec- tional preferences. We compare three variants of noun class information: (i) classes induced from the lexical resource GermaNet or obtained from clusterings based on either (ii) window information or (iii) syntactic features. Furthermore, we experiment with PP rule generaliza- tion. While we do not significantly improve over the baseline, our results demonstrate that (i) integrating selectional preferences as rigid class annotation in the parse tree is sub-optimal, and that (ii) clusterings based on window co-occurrence are more robust than syntax-based clusters or GermaNet classes for the task of modeling selectional preferences. 1 Introduction The translation of prepositions is difficult in SMT: some prepositions convey a meaning (to sit UNDER/ON the table) while others are merely functional (to believe IN sth.). Both kinds of prepositions pose a significant challenge to the translation system as they largely depend on target-language-specific restrictions for which there is often not enough contextual information available. Translating prepositions is difficult as SMT systems must choose the correct preposi- tion given the intended meaning of the preposition in the input sentence as well as the target-side context in which the preposition appears. In English-to-German translation, there are cases in which the target-language does not require a preposition (e.g. to call FOR → fordern), or in which it is necessary to produce a target-language preposition even though there is no preposition in the input sentence (e.g. to enter → gelangen IN). The choice of prepositions is typically determined by a governor, such as verbs (to believe in sth.) or nouns (e.g. interest in sth.). In addition, the preposition depends on the semantic class of nouns that are governed. For example, to learn from can lead to two different translations in German: in the case of to learn from [a person], the correct translation is lernen VON, whereas to learn from [the past] should be translated as lernen AUS. We present a novel method that uses noun class information to model selectional prefer- ences of prepositions. By annotating noun class information into the parse trees used to train an English-German string-to-tree SMT-system, we aim at obtaining more precise translation rules. Instead of allowing any PP in a given rule, the noun class annotation restricts that rule to PPs of a specific semantic class. While this procedure adds semantically fine-grained information, Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC © The Authors 275 it also leads to a loss of rule generalization. We compensate for this loss by making generic, non-annotated rules accessible for the enriched system, and by generating new PP rules that cannot be derived from the parallel data. The selectional …

Content truncated to 500 words